OpenAI says its AI is getting better at human everyday tasks like shopping and meal planning

OpenAI said Monday that it has built a new internal tool for measuring how well its models handle the kind of everyday errands people already bring to ChatGPT, from assembling a grocery cart to filling out online forms and creating a budget, and released research on how its current models performed on it.

The headline numbers come from tasks designed to look like real life. According to CNN, GPT-6 Astra, which OpenAI released in September, scored 93.1% at finding the right kindergarten after a move, 98% at managing community volunteers preparing for a flood, and 90.4% at creating a meal plan for a parent recently diagnosed with diabetes.

On a broader measure covering everyday-task prompts, GPT-6 Astra scored 80%. The comparison point is GPT-5.2, the model OpenAI shipped in December 2025, which scored 49.3% — a gap that frames the past year as a jump in practical usefulness rather than a rounding error on a leaderboard.

The evaluation targets the skills a model needs to get these answers right: navigating websites, applying context pulled from large amounts of chat history, and reaching conclusions that match a user's needs and expectations. The tasks themselves range widely, from apartment hunting to caring for a loved one.

OpenAI's stated motive is visibility. Because so many people use the app, the company says it wants to capture the full range of use cases, track how performance changes over time, and identify what the models are still not good at.

Grading work like this is harder than scoring a math test. CNN reports that the evaluation relies on people with relevant expertise to establish what a correct answer looks like, since whether a meal plan or a volunteer schedule is right depends on the specific family, the specific flood, and other context that has no single answer.

The results land as OpenAI, Meta, Anthropic and Google race to build agents that can complete real-world tasks, while all of them face safety scrutiny and pressure to turn massive infrastructure spending into revenue. A number that says a model can handle errands is useful currency in that contest.

The obvious caveat is that these are company-reported scores from an internal tool, and outside researchers cannot yet check them. What to watch next: whether OpenAI publishes the evaluation details or the tool itself, and how quickly scores on everyday chores turn into shipped agent features inside ChatGPT.